Workers typically perform a variety jobs over the course of their careers in firms. These jobs differ in the skills they require and, thus, may convey different amounts of information about worker ability. In this paper, we develop a theory of learning and job design, and empirically test this theory using a unique firm-level data set. Our model departs from the existing literature on learning and sorting by recognizing that the information content of jobs may differ and that optimizing firms will consider the information they gather about workers when assigning them to different jobs. In addition, we recognize that jobs are mutable and that firms may alter job definitions in order to efficiently learn about workers and make prudent decisions about whom to promote. As a result, the extent to which on-the-job performance reveals worker ability is determined endogenously. A key implication of the model is that firms may not assign workers to the jobs at which they are expected to be the most productive, particularly when they are less experienced. The reason is that some jobs provide more accurate information on skills than do others, and this information is valuable in making future job assignments. Thus, there may be an option value to assigning workers to jobs in which their expected performance is relatively low. A second key implication is that career paths are history-dependent. Workers with relatively low performance may be reassigned to jobs which are less sensitive to talent and, thus, convey little information about worker ability. Thus, current productivity shocks can have lasting effects on a worker's career, and there are conditions in which good workers can become (permanently) trapped in bad jobs. Job assignment is, therefore, inefficient (compared to a full information benchmark) even when the time horizon is sufficiently long. In contrast to the assumptions underlying multi-armed bandit models of job mobility, in our model, a worker's ability in one job is not independent of his or her ability in other jobs. Thus, standard solution techniques (see Gittins and Jones, 1974) do not apply. Instead, we frame our model as a dynamic programming problem and solve it numerically. We then characterize the patterns of job assignments, job duration and job transitions over a worker's career and relate these patterns to the firm's prior beliefs about worker productivity and to on-the-job performance. We then test the empirical predictions of our theory using 20 years of personnel data from a single firm. Initial results provide tentative support for the model's predictions (we thank George Baker for generously providing us with the data)
